Development of the Neutrophil-to-Platelet Ratio (NPR) Integrated with Machine Learning for Predicting Early Mortality After Mechanical Thrombectomy in Acute Ischemic Stroke

Abstract

Purpose

This study aimed to evaluate the predictive value of inflammatory markers, particularly the neutrophil-to-platelet ratio (NPR), combined with clinical parameters for early mortality following mechanical thrombectomy (MT) in patients with large artery occlusive acute ischemic stroke (LAO-AIS), to guide timely clinical interventions.

Patients and methods

This retrospective study analyzed 320 LAO-AIS patients who underwent MT between January 2023 and January 2025. Missing data (<15%) were imputed. Boruta feature selection identified variables for multiple logistic regression. The dataset was randomly divided into training and test sets (7:3). A nomogram was constructed, and four machine learning algorithms—Decision Tree (DT), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Naive Bayes (NB)—were developed and validated.

Results

Early mortality occurred in 67 cases. Multivariate analysis identified six independent predictors: standardized NPR (NPR_std; OR = 4.51, P < 0.001), age (OR = 1.10, P < 0.001), decompressive craniectomy (DC; OR = 0.19, P < 0.001), responsible artery location (OR = 0.34, P = 0.006), lymphocyte count (LYM; OR = 2.14, P = 0.008), and prothrombin time (PT; OR = 1.31, P = 0.011). The nomogram showed high reliability. XGBoost achieved superior predictive performance, with SHapley Additive exPlanations (SHAP) analysis confirming NPR_std as the most important predictor.

Conclusion

The neutrophil-to-platelet ratio (NPR) is an independent predictor of early mortality after MT in LAO-AIS patients. The predictive model provides valuable guidance for clinicians to adjust treatment strategies early.

Introduction

Stroke ranks as the second leading cause of global mortality and third in Disability-Adjusted Life Years (DALYs), projected to ascend to the second position by 2050.¹ Ischemic stroke predominates (65.3% of incident cases), with large artery occlusion (LAO) constituting 30–40% of acute ischemic stroke (AIS) and strongly correlating with 90-day mortality.^2,3 MT is established as the standard-of-care for LAO-AIS through pivotal randomized trials.^4–6 Nevertheless, substantial disparities persist: 90-day mortality and disability rates remain markedly elevated in low- and middle-income countries (LMICs), which bear 83.3% of the global stroke burden, with Asia contributing 61%.^7–10 Despite MT’s efficacy, reperfusion injury and malignant cerebral edema represent prevalent, life-threatening complications that significantly contribute to post-procedural mortality, underscoring the need for improved risk stratification tools.¹¹

Post-ischemic stroke, rapid neuroinflammation occurs at the infarction core.¹² Neutrophils, as pivotal mediators in the thromboinflammatory cascade, critically interact with platelets, amplifying inflammation and dysregulating the immune microenvironment – processes directly implicated in secondary brain injury.^13,14 The NPR, calculated from admission peripheral blood, reflects a relative neutrophilia and systemic inflammatory state. Elevated NPR signifies enhanced neutrophil activation and platelet consumption, facilitating neutrophil recruitment to the infarct via thromboinflammatory pathways and establishing a deleterious “vicious cycle” that exacerbates blood-brain barrier disruption and edema.¹⁴ Critically, excessive neuroinflammation driven by these mechanisms (eg, complement activation, inflammatory mediator release) worsens tissue damage, infarct expansion, and poor outcomes.¹⁵ Given evidence that early anti-inflammatory intervention improves prognosis,^16,17 this study aims to develop and validate a novel NPR-based predictive model, leveraging machine learning, to forecast early (90-day) post-MT mortality in LAO patients. This will enable timely identification of high-risk individuals for targeted therapeutic strategies.

Methods

Patient Selection Statistical Analysis

This retrospective study analyzed data from all 320 patients diagnosed with LAO-AIS who underwent MT at the Affiliated Hospital of Guilin Medical University between January 2023 and January 2025. Data with a missing rate of ≥15% were excluded, while those with a missing rate of ≤15% were processed using multiple imputation. The study was approved by the Ethics Committee of the Affiliated Hospital of Guilin Medical University and conducted in accordance with the Declaration of Helsinki.

Patients were included if they met the following criteria (Figure 1):

Figure 1.

Patient selection.

Abbreviations: LAO-AIS, large artery occlusion acute ischemic stroke; MT, mechanical thrombectomy.

1. Admission within 24 hours of symptom onset, presenting with neurological deficits confirmed as LAO-AIS by cranial CT angiography (CTA) or digital subtraction angiography (DSA).

2. Eligibility for MT and subsequent treatment at our hospital. The indications for mechanical thrombectomy included:

NIHSS score ≥ 6 with symptoms attributable to the occlusion of the responsible vessel.

Alberta Stroke Program Early CT Score (ASPECTS) ≥ 6; for patients with large infarctions and ASPECTS ≤ 5, thrombectomy was performed if a salvageable ischemic penumbra was present.¹⁸

Absence of intracranial hypodense lesions affecting more than one-third of the cerebral hemisphere or intracranial hemorrhage as confirmed by CT or Magnetic Resonance Imaging (MRI).

For anterior circulation strokes, imaging criteria met the Deferring Antiplatelet Therapy Until Symptomatic Intracranial Stenosis (DAWN) or Diffusion and Perfusion Imaging Evaluation for Understanding Stroke Evolution with Intravenous Alteplase (DEFUSE-3) guidelines, allowing an extended treatment window of up to 24 hours post-onset.

For posterior circulation strokes due to basilar artery occlusion, the MT window was extended to 12–24 hours based on the Basilar Artery Occlusion Chinese Endovascular Trial (BAOCHE trial).¹⁹

3. Age ≥ 18 years.

Patients were excluded if they met the following criteria:

(1) Lack of blood test information, n=14 (2) Comorbid Life-Threatening Conditions: Patients diagnosed with severe comorbidities that carry a high short-term mortality risk, which could independently influence the study outcome and confound data interpretation. Specific exclusions in this category encompassed. Severe Liver Disease: Such as decompensated cirrhosis or severe hepatitis. Active Malignancy: Both treated and untreated metastatic or hematological cancers. End-Stage Organ Failure: Including end-stage renal disease (on renal replacement therapy), severe heart failure (NYHA Class III–IV), and advanced respiratory failure (eg, requiring long-term oxygen therapy).(Total number of patients excluded based on the above conditions: n=11) (3) Patients with venous vessel occlusion, n=8 (4) Lost to follow-up patients, n=15

Early mortality was defined as death occurring at discharge or within 90 days post-discharge.

Data Collection

The research team gathered baseline clinical data for all patients, encompassing age, gender, atrial fibrillation, hyperlipidemia, hypertension, diabetes, smoking and alcohol consumption. Mortality data were ascertained through medical records for inpatients and follow-up visits/telephone interviews for discharged patients. LAO and responsible arterywas defined as occlusion of proximal intracranial arteries, including: Anterior circulation: Internal carotid artery (cervical/intracranial), M1/proximal M2 segment of middle cerebral artery, A1 segment of anterior cerebral artery; Posterior circulation: Intracranial vertebral artery (V4), basilar artery, P1 segment of posterior cerebral artery. Occlusion was confirmed by CTA/MRA/DSA with vessel diameter ≥2.0 mm. Hyperlipidemia is characterized by elevated triglycerides and/or total cholesterol, elevated Low-Density Lipoprotein Cholesterol (LDL-C), and reduced High-Density Lipoprotein Cholesterol (HDL-C). Hypertension was defined as a documented history of the condition at admission or a systolic/diastolic blood pressure of ≥140/90 mmHg upon admission. Diabetes was identified based on a prior diagnosis at admission. Atrial fibrillation was classified as either a pre-existing condition or a diagnosis confirmed by electrocardiogram. The study utilized the first laboratory test results obtained at admission. The original baseline data are detailed in Appendix A. Early mortality was defined as death occurring either at discharge or within 90 days post-discharge. The NPR was calculated as NPR = neutrophils/platelets. Given that the raw NPR values might be excessively small, potentially compromising the model’s stability, the study applied centering and standardization to NPR, represented as NPR_std.

Statistical Analysis

Continuous baseline data following a normal distribution were presented as mean ± standard deviation, with intergroup comparisons performed using t-tests. For non-normally distributed continuous data, values were expressed as median and interquartile range, and intergroup comparisons were conducted using rank sum tests. Categorical variables were reported as percentages, n (%), with comparisons between groups performed using chi-square tests. When more than 20% of the cells in the contingency table had expected frequencies of <5, Fisher’s exact test was applied. Multiple imputation, implemented via the mice package, was used to address missing data comprising less than 15% of the original dataset, with 10 imputations performed. The dataset with the lowest Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values was selected for analysis in this study (Appendix B).²⁰

The Boruta feature selection algorithm was employed to identify variables for inclusion in the logistic multiple regression analysis, with 100 iterations.²¹ Independent predictors were then integrated to construct a predictive model, which was randomly partitioned into a training set (n = 224) and an internal validation set (n = 96) at a 7:3 ratio. The model’s predictive performance was assessed using the receiver operating characteristic (ROC) curve, while its discriminative capability was verified through calibration plots. Decision curve analysis was conducted to evaluate the practical utility of the nomogram in clinical decision-making. Additionally, four machine learning algorithms—DT, XGBoost, SVM, and NB—were employed for model development. Among these, the best-performing XGBoost model underwent SHAP analysis to interpret variable importance. All statistical analyses were conducted using R software (version 4.4.2), with a significance threshold of P < 0.05.

Results

All results were derived from the multiply imputed dataset. The study included a total of 320 participants, comprising 67 deceased and 253 surviving patients. Significant baseline differences were observed in Age, DC, Responsible Artery, WBC, NE, PLT, NPR, and NPR_std (P < 0.05; Table 1). The baseline characteristics of the imputed dataset showed no substantial deviation from those of the original data (Appendix A).

Table 1.

Patient Baseline Data Table

	ALL Patients	Alive	Dead	P
	N=320	N=253	N=67
Age	56.00 (52.00, 65.00)	55.00 (51.00, 63.00)	66.00 (56.50, 75.00)	<0.001
Gender				0.304
Male	215 (67%)	174 (68%)	41 (61%)
Female	105 (33%)	79 (32%)	26 (39%)
DC				0.001
No	180 (56%)	130 (51%)	50 (75%)
Yes	140 (44%)	123 (49%)	17 (25%)
Responsible artery				<0.001
Posterior circulation	133 (42%)	90 (36%)	43 (64%)
Anterior circulation	187 (58%)	163 (64%)	24 (36%)
Hypertension				0.833
No	123 (38%)	96 (38%)	27 (40%)
Yes	197 (62%)	157 (62%)	40 (60%)
Atrial fibrillation				0.071
No	244 (76%)	199 (79%)	45 (67%)
Yes	76 (24%)	54 (21%)	22 (33%)
Hyperlipidemia				0.117
No	247 (77%)	190 (75%)	57 (85%)
Yes	73 (23%)	63 (25%)	10 (15%)
Diabetes				0.841
No	230 (72%)	183 (72%)	47 (70%)
Yes	90 (28%)	70 (28%)	20 (30%)
Smoking				0.631
No	190 (59%)	148 (58%)	42 (63%)
Yes	130 (41%)	105 (42%)	25 (37%)
Alcohol				0.998
No	246 (77%)	195 (77%)	51 (76%)
Yes	74 (23%)	58 (23%)	16 (24%)
SBP	146.00 (129.75, 163.00)	147.00 (130.00, 164.00)	142.00 (129.00, 158.50)	0.339
DBP	85.50 (77.00, 96.00)	87.00 (77.00, 97.00)	84.00 (75.50, 93.50)	0.260
WBC	12.65 (9.90, 15.20)	12.20 (9.80, 14.80)	13.40 (10.80, 15.95)	0.012
NE	9.64 (7.57, 12.16)	9.55 (7.35, 12.00)	10.50 (8.31, 12.75)	0.045
PLT	287.50 (196.00, 361.25)	306.00 (215.00, 380.00)	196.00 (157.00, 278.50)	<0.001
NPR	0.04 (0.03,0.05)	0.03 (0.02,0.04)	0.05 (0.04,0.07)	<0.001
NPR_std	0.08 (−0.53,0.70)	−0.53 (−1.03,0.09)	0.71 (0.09,1.93)	<0.001
LYM	1.28 (0.94,1.86)	1.24 (0.90,1.90)	1.40 (1.08,1.77)	0.275
MO	0.79 (0.56,1.01)	0.77 (0.55,1.00)	0.82 (0.64,1.13)	0.175
PT	13.80 (13.00,15.20)	13.80 (13.00,14.80)	13.90 (12.90,16.05)	0.295
APTT	35.50 (32.60,38.80)	35.50 (32.60,39.00)	35.50 (32.50,38.20)	0.796
D-D	1.01 (0.55,1.65)	1.03 (0.53,1.64)	0.88 (0.59,1.69)	0.839
FIB	3.36 (2.80,4.10)	3.36 (2.80,3.94)	3.35 (2.89,4.22)	0.266
LDL	2.77 (2.17,3.42)	2.78 (2.18,3.41)	2.63 (2.10,3.49)	0.642
HDL	1.06 (0.92,1.26)	1.06 (0.92,1.25)	1.09 (0.92,1.33)	0.453
ALB	39.80 (36.80,42.20)	40.00 (36.90,42.20)	39.10 (36.65,42.20)	0.596
K	3.64 (3.41,3.87)	3.67 (3.41,3.88)	3.60 (3.40,3.80)	0.208
Ca	2.18 (2.11,2.27)	2.18 (2.10,2.27)	2.20 (2.12,2.25)	0.629
Cl	103.00 (101.00,105.00)	103.00 (101.00,105.00)	103.00 (100.50,105.00)	0.852

Notes: Significant baseline differences were observed in Age, DC, Responsible Artery, WBC, NE, PLT, NPR, and NPR_std (P < 0.05).

Abbreviations: DC, decompressive craniectomy; SBP, systolic blood pressure; DBP, diastolic blood pressure; WBC, white blood cell; NE, neutrophil granulocyte; PLT, platelet; NPR, neutrophil-to-platelet ratio; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; MO, monocyte; PT, prothrombin time; APTT, activated partial thromboplastin time; D-D, d-dimer; FIB, fibrinogen; LDL, low-density lipoprotein; HDL, high-density lipoprotein; ALB, albumin; K, potassium; Ca, calcium; Cl, chlorine.

Eight variables, namely NPR_std, Age, DC, WBC, Responsible arterie, LYM, PT, and D-D, were selected by Boruta feature screening (Figure 2). As NPR_std is composed of neutrophils and platelets, these two variables were not incorporated into the Boruta feature screening. The feature details are presented in Appendix C.

Figure 2.

Boruta feature screening map.

Notes: The core of Boruta is to test whether a feature’s predictive importance is significantly greater than that of randomized “shadow features” created by permuting the original data. This establishes a benchmark for significance (the maximum shadow feature importance, or “shadow Max,” represented by the rightmost blue marker). Through an iterative process, features were categorized as follows: Confirmed predictors (Green), including NPR_std, Age, DC, etc., whose importance consistently surpassed the “shadow Max” threshold; and Rejected features (Red), which were deemed statistically insignificant as their importance was lower than the random benchmark. This rigorous iterative testing guarantees the selection of features that demonstrate stable and significant importance beyond random chance.

Abbreviations: DC, decompressive craniectomy; SBP, systolic blood pressure; DBP, diastolic blood pressure; WBC, white blood cell; NPR_std, centralized and standardized neutrophil-to-platelet ratio; LYM, lymphocyte; MO, monocyte; PT, prothrombin time; APTT, activated partial thromboplastin time; D-D, d-dimer; FIB, fibrinogen; LDL, low-density lipoprotein; HDL, high-density lipoprotein; ALB, albumin; K, potassium; Ca, calcium; Cl, chlorine.

The selected variables were incorporated into the logistic multiple regression analysis (Figure 3). As shown in Table 2 and Figure 3, NPR_std, Age, DC, Responsible Artery, LYM, and PT were identified as independent predictors of mortality in LAO-AIS patients. The results indicate that for each additional year of age, the risk of death increases by 0.10 times (P < 0.001, OR = 1.10, 95% CI = 1.05–1.15). Patients who underwent DC exhibited a 0.80-fold reduction in mortality risk (P < 0.001, OR = 0.19, 95% CI = 0.07–0.44). Each unit increase in NPR_std was associated with a 3.51-fold increase in mortality risk (P < 0.001, OR = 4.51, 95% CI = 2.90–7.49). Compared to patients with posterior circulation involvement, those with anterior circulation as the responsible artery had a 0.65-fold decrease in mortality risk (P = 0.006, OR = 0.34, 95% CI = 0.15–0.73). Each unit increase in LYM level was linked to a 1.143-fold increase in mortality risk (P = 0.008, OR = 2.14, 95% CI = 1.25–3.89). Furthermore, for each additional second of PT, the risk of death increased by 0.31 times (P = 0.002, OR = 1.31, 95% CI = 1.06–1.62).

Figure 3.

Multiple regression forest plot.

Notes: Only independent predictive factors were included in the figure, while variables without statistical significance in multiple regression were excluded. Age, NPR_std, LYM, and PT were identified as independent risk factors (OR > 1), whereas DC and the responsible artery served as independent protective factors (OR < 1).

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time.

Table 2.

Multivariate Logistic Regression Analysis

Characteristics	B	OR (95% CI)	P
Age	0.09	1.10 (1.05–1.15)	<0.001
DC	−1.62	0.19 (0.07–0.44)	<0.001
NPR_std	1.50	4.51 (2.90–7.49)	<0.001
Responsible artery	−1.06	0.34 (0.15–0.73)	0.006
LYM	0.76	2.14 (1.25–3.89)	0.008
PT	0.27	1.31 (1.06–1.62)	0.011
WBC	−0.02	0.97 (0.88–1.07)	0.652
D-D	−0.20	0.81 (0.48–1.32)	0.41

Notes: NPR_std, Age, DC, Responsible Artery, LYM, and PT are independent predictive factors (P < 0.05). Among them, Age, NPR_std, LYM, and PT are independent risk factors (B > 0, OR > 1), while DC and Responsible Artery serve as independent protective factors (B < 0, OR < 1). WBC and D-D did not demonstrate statistical significance (P > 0.05).

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time; WBC, white blood cell; D-D, d-dimer.

In further regression and stratified analysis, Table 3 indicates that within the investigated patient population, older age (≥56 years) and higher NPR_std (≥0.075) are independent risk factors for the occurrence of the primary outcome. Conversely, the presence of DC (YES) and anterior circulation (AC) as the responsible artery are independent protective factors against the occurrence of the primary outcome. Lymphocyte count (LYM) and prothrombin time (PT) were not found to be significantly associated with the outcome in this study.

Table 3.

Independent Logistic Regression and Stratified Analyses

Characteristics	Patients	Univariate Analysis		Multivariate Analysis
		OR (95% CI)	P	OR (95% CI)	P
Age			<0.001		0.005
<56	173	1		1
≥56	147	4.19(2.21,7.94)		2.80(1.37,5.7)
DC			<0.001		<0.001
No	180	1		1
Yes	140	0.35(0.19,0.65)		0.26(0.13,0.53)
NPR_std			<0.001		<0.001
<0.075	160	1		1
≥0.075	160	8.30(4.05,17.00)		9.4(4.34,20.39)
Responsible artery			<0.001		<0.001
PC	133	1		1
AC	187	0.31(0.17,0.54)		0.33(0.17,0.64)
LYM			0.337		0.289
<1.28	160	1		1
≥1.28	160	1.30(0.75,2.23)		1.42(0.74,2.73)
PT			0.865		0.649
<13.8	151	1		1
≥13.8	169	1.04(0.61,1.79)		1.16(0.6,2.24)

Notes: NPR_std, Age, DC and Responsible Artery are independent predictive factors (P < 0.05).

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time; WBC, white blood cell; D-D, d-dimer.

There were statistically significant differences in NPR_std, Age, DC, and the responsible artery between the survival group and the death group. There is no statistical difference between LYM and PT (Figure 4).

Figure 4.

Differential analysis between survival group and death group.

Notes: (A–F) show the distribution of NPR_std, Age, LYM, PT, responsible artery, and DC in the alive versus dead groups, respectively. There were statistically significant differences in NPR_std, Age, DC, and the responsible artery between the survival group and the death group, Asterisks denote statistical significance: **** for P < 0.0001, *** for P < 0.001, and ns for non-significant (P > 0.05).

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time; PC, posterior circulation; AC, anterior circulation.

ROC curve analysis (Figure 5) showed that NPR_std exhibited the best predictive performance with an AUC of 0.794 (P<0.001), followed by Age with an AUC of 0.759 (P<0.001). Resp. Art. and DC also demonstrated statistically significant predictive value, with AUCs of 0.643 (P<0.001) and 0.616 (P=0.003), respectively. In contrast, LYM (AUC=0.543, P=0.275) and PT (AUC=0.541, P=0.294) showed poor discriminative ability, as their AUCs were close to 0.5 and the P-values were greater than 0.05, indicating no significant predictive efficacy for the target outcome.

Figure 5.

Roc analysis.

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time; PC, posterior circulation; AC, anterior circulation; Resp. Art., responsible artery.

The graph illustrates the association between standardized NPR (NPR std) and an outcome variable (Figure 6). Statistical analysis reveals a highly significant overall association (P < 0.001), indicating a notable correlation between changes in NPR std and the outcome variable, while the non-linear test yields a P-value of 0.852 (P > 0.05), suggesting no significant evidence of a non-linear relationship.

Figure 6.

Analysis of RCS for NPR_std.

Table 4 shows the data of the randomly split training set and test set, indicating that the two groups of data are basically balanced and comparable.

Table 4.

Baseline Tables for Training and Testing Sets

	ALL Patients	Test	Train	P
	N=320	N=96	N=224
Age	56.00 (52.00,65.00)	56.00 (50.75,65.00)	57.00 (52.00,65.25)	0.193
Gender				0.194
Male	215 (68%)	59 (61%)	156 (70%)
Female	105 (32%)	37 (39%)	68 (30%)
DC				0.037
No	180 (56%)	45 (47%)	135 (60%)
Yes	140 (44%)	51 (53%)	89 (40%)
Dead				0.905
No	253 (79%)	75 (78%)	178 (79%)
Yes	67 (21%)	21 (22%)	46 (21%)
Responsible artery				0.729
Posterior circulation	133 (42%)	38 (40%)	95 (42%)
Anterior circulation	187 (58%)	58 (60%)	129 (58%)
Hypertension				0.688
No	123 (38%)	39 (41%)	84 (38%)
Yes	197 (62%)	57 (59%)	140 (62%)
Atrial fibrillation				0.626
No	244 (76%)	71 (74%)	173 (77%)
Yes	76 (24%)	25 (26%)	51 (23%)
Hyperlipidemia				0.295
No	247 (77%)	70 (73%)	177 (79%)
Yes	73 (23%)	26 (27%)	47 (21%)
Diabetes				0.684
No	230 (72%)	67 (70%)	163 (73%)
Yes	90 (28%)	29 (30%)	61 (27%)
Smoking				0.709
No	190 (59%)	59 (61%)	131 (58%)
Yes	130 (41%)	37 (39%)	93 (42%)
Alcohol				0.707
No	246 (77%)	72 (75%)	174 (78%)
Yes	74 (23%)	24 (25%)	50 (22%)
SBP	146.00 (129.75,163.00)	143.50 (127.50,160.00)	147.50 (131.00,164.00)	0.216
DBP	85.50 (77.00,96.00)	85.00 (76.75,93.00)	86.00 (77.00,97.00)	0.316
WBC	12.65 (9.90,15.20)	13.00(11.02,15.32)	12.35 (9.77,15.20)	0.147
NE	9.64 (7.57,12.15)	10.29 (8.31,12.18)	9.57 (7.35,12.06)	0.126
PLT	287.50 (196.00,361.25)	306.00 (237.50,369.25)	276.00 (189.50,356.25)	0.054
NPR	0.04 (0.03,0.05)	0.03 (0.03,0.04)	0.04 (0.03,0.05)	0.609
NPR_std	0.07 (−0.53,0.69)	−0.52 (−0.53,0.22)	0.076(−0.53,0.70)	0.609
LYM	1.28 (0.94,1.86)	1.350 (1.01,2.02)	1.230 (0.91,1.79)	0.171
MO	0.79 (0.55,1.01)	0.86 (0.62,1.13)	0.76(0.55,0.97)	0.045
PT	13.80 (13.00,15.20)	14.10 (13.07,15.25)	13.70 (12.90,15.20)	0.280
APTT	35.50 (32.60,38.80)	35.45 (32.77,37.60)	35.60 (32.27,39.00)	0.828
D-D	1.01 (0.54,1.64)	1.01 (0.69,1.47)	1.01 (0.53,1.70)	0.754
FIB	3.36 (2.80,4.10)	3.35 (2.77,4.07)	3.38 (2.87,4.10)	0.548
LDL	2.77 (2.17,3.42)	2.75 (2.13,3.40)	2.77 (2.18,3.44)	0.889
HDL	1.06 (0.92,1.26)	1.01 (0.90,1.21)	1.08 (0.93,1.27)	0.182
ALB	39.80 (36.80,42.20)	38.85 (36.80,41.72)	40.10 (36.80,42.30)	0.162
K	3.64 (3.41,3.87)	3.66 (3.44,3.86)	3.63 (3.39,3.87)	0.680
Ca	2.18 (2.11,2.27)	2.20 (2.12,2.27)	2.18 (2.10,2.26)	0.251
Cl	103.00 (101.00,105.00)	103.00 (101.75,105.00)	103.00 (100.75,105.00)	0.455

Notes: There was no significant statistical difference in baseline characteristics between the training and test sets (P > 0.05), except for DC (P = 0.039) and MO (P = 0.045).

Abbreviations: DC, decompressive craniectomy; SBP, systolic blood pressure; DBP, diastolic blood pressure; WBC, white blood cell; NE, neutrophil granulocyte; PLT, platelet; NPR, neutrophil-to-platelet ratio; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; MO, monocyte; PT, prothrombin time; APTT, activated partial thromboplastin time; D-D, d-dimer; FIB, fibrinogen; LDL, low-density lipoprotein; HDL, high-density lipoprotein; ALB, albumin; K, potassium; Ca, calcium; Cl, chlorine.

This study identified six independent predictive factors: NPR_std, Age, DC, Responsible Artery, LYM, and PT. A numerical value was assigned to each factor, and the total score was derived by summing these individual scores. This composite score served as an indicator for predicting early mortality risk in LAO-AIS patients (Figure 7).

Figure 7.

A nomogram for predicting early death in patients with Lao - AIS.

Notes: The sixth sequentially sorted patient in the input dataset was selected as an illustrative case, demonstrating an early mortality probability of 0.405.

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of the neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time; PC, Posterior circulation; AC, Anterior circulation.

In the predictive model developed in this study, the AUC for the training set in logistic regression was 0.926 (95% CI = 0.885–0.967), while the AUC for the validation set was 0.853 (95% CI = 0.753–0.952), demonstrating strong predictive performance (Figure 8A). To ensure the reliability of the calibration curve analysis, 500 internal repeated samplings were conducted for both the training and validation sets. The results exhibited a high degree of concordance between the predicted and observed values in the nomogram (Figure 8BI and Bii). Additionally, Figure 8Ci, 8Cii, 8Di, and 8Dii illustrate that clinical interventions guided by this prediction model provided substantial clinical benefits.

Figure 8.

ROC curve, calibration curve and DCA plot.

Notes: The AUC of the training and test sets in (A), (Bi and Bii) are the calibration curves of the internally repeated sampling training and test sets, respectively. (Ci and Di) are the Decision Curve Analysis of the training set, and (Cii and Dii) are the Decision Curve Analysis of the test set.

To identify the optimal predictive model, this study employed and rigorously evaluated four distinct machine learning algorithms: Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Naïve Bayes (NB). The models were assessed using a comprehensive set of metrics, each selected for its specific insight into model behavior. Accuracy provides an overall view of correctness, while Sensitivity (Recall) and Specificity are critical in clinical settings for quantifying the model’s ability to identify true positives and true negatives, respectively. The F1-Score, as the harmonic mean of Precision and Recall, offers a single metric to balance the concern of missing true cases (low Recall) against the cost of false alarms (low Precision). Finally, the Brier Score evaluates the calibration of predicted probabilities, where a lower score indicates more reliable and confident predictions.

Based on the results (Table 5 and Figure 9), a detailed comparison reveals distinct strengths and weaknesses. On the training set, XGBoost achieved near-perfect sensitivity (0.95) and the highest accuracy (0.89), demonstrating its powerful capacity for pattern recognition. However, its performance on the validation set provides the true test of generalizability. Here, while the tree-based DT model exhibited the highest F1-score (0.66) and competitive accuracy (0.81), its significant drop in sensitivity from train to test (0.69 to 0.72) compared to XGBoost’s more pronounced drop (0.95 to 0.66) suggests that DT may be less complex and thus more robust, whereas XGBoost shows signs of potential overfitting to the training data. This is further supported by the Brier Score; although all models had identical test set Brier Scores (0.11), XGBoost’s notably lower training Brier Score (0.06) confirms its superior probability calibration on seen data.

Table 5.

Machine Learning Feature Table

Model	Dfclass	Sensitivity	Specificity	Precision	Recall	F1	Prevalence	Accuracy	Brier Score
LR	Train	0.89	0.83	0.57	0.89	0.70	0.20	0.86	0.078(0.05–0.10)
DT	Train	0.69	0.94	0.78	0.69	0.73	0.23	0.81	0.086(0.06–0.11)
XGB	Train	0.95	0.82	0.58	0.95	0.72	0.20	0.89	0.06(0.04–0.08)
SVM	Train	0.82	0.88	0.64	0.82	0.72	0.20	0.85	0.08 (0.06–0.10)
NB	Train	0.89	0.79	0.53	0.89	0.66	0.20	0.84	0.08(0.05–0.11)
LR	Test	0.61	0.84	0.52	0.61	0.56	0.21	0.73	0.11(0.07–0.18)
DT	Test	0.72	0.89	0.61	0.72	0.66	0.18	0.81	0.11(0.07–0.17)
XGB	Test	0.66	0.85	0.56	0.66	0.60	0.21	0.76	0.11(0.07–0.16)
SVM	Test	0.61	0.85	0.54	0.61	0.57	0.21	0.73	0.11(0.07–0.17)
NB	Test	0.71	0.88	0.62	0.71	0.66	0.21	0.79	0.11(0.07–0.18)

Notes: In the training set, the model with the highest AUC is the XGBoost model (0.95), the one with the highest Precision is the DT model (0.78), the one with the highest Recall is the XGBoost model (0.95), the one with the highest F1 is the DT model (0.73), and the one with the highest Accuracy is the XGBoost model (0.89). The model with the lowest Brier Score is the XGBoost model (0.08). In the test set, the model with the highest AUC is the NB model (0.86), the one with the highest Precision is the NB model (0.62), the one with the highest Recall is the DT model (0.72), the one with the highest F1 is both the DT and NB models (0.66), the one with the highest Accuracy is the DT model (0.81), and the model with the lowest Brier Score is the XGBoost model (0.11).

Abbreviations: LR, logistic regression; DT, decision tree; XGBoost, extreme gradient boosting; SVM, support vector machine; NB, naive bayes; Dfclass, different classes.

Figure 9.

Machine learning feature maps.

Notes: (Ai and Bi) are the machine learning results of the training set, while (Aii and Bii) are the machine learning results of the test set.

Abbreviations: LR, logistic regression; DT, decision tree; XGBoost, extreme gradient boosting; SVM, support vector machine; NB, naive bayes; Dfclass, different classes.

Therefore, the selection of the XGBoost model as the final predictor is a nuanced decision. It is justified not merely by a single metric but by a holistic view of its high discriminative ability (as evidenced by the leading validation AUC of 0.86), its excellent calibration, and its best-in-class performance on the training data which suggests high potential with a larger dataset. While it may be more complex, its overall performance profile makes it the most suitable model for the predictive task at hand.

The SHAP analysis of the model developed using XGBoost reveals that NPR_std holds the highest importance in this model (Figure 10A). In the SHAP value plots of each variable, NPR_std, Age, and LYM exhibit an approximately positive correlation, while PT takes the form of a downward-opening curve (Figure 10B). We also present the SHAP value plot of one of the patients (Figure 10C).

Figure 10.

SHAP Explanation Diagram.

Notes: (A) represents the SHAP analysis of all variables, (B) depicts the distribution of SHAP values for individual variables, and (C) shows the SHAP values for a particular case.

Abbreviations: DC, decompressive craniectomy; NPR_std, centralized and standardized values of neutrophil-to-platelet ratio; LYM, lymphocyte; PT, prothrombin time.

Discussion

This study primarily investigates the predictive factors for mortality within 90 days following mechanical thrombectomy in patients with LAO-AIS and develops a nomogram centered on the inflammatory marker NPR. The significance of NPR is further validated through four machine learning approaches.

Despite growing focus on post-MT prognosis in LAO-AIS, critical gaps persist in existing predictive approaches. Most rely on indirect inflammatory indices (eg, NLR, PLR)²² that fail to directly capture neutrophil-platelet thrombotic-inflammatory crosstalk—key to MT-induced reperfusion injury.²³ Unstandardized cell ratios and overreliance on simple models (eg, logistic regression alone) further limit stability and discriminative power, while few tools are tailored to MT’s unique pathophysiology (eg, recanalization-triggered neuroinflammation), rather than unselected AIS.²⁴

To address these, our study offers targeted innovations: (1) NPR_std (centralized/standardized NPR) eliminates scaling biases to reflect neutrophil-platelet interactions accurately; (2) Boruta feature selection paired with XGBoost (outperforming traditional models) enhances predictive robustness; (3) a concise nomogram integrates NPR_std with clinical factors for bedside utility. This fills the unmet need for MT-specific, inflammation-centered risk tools, aiding timely interventions to reduce early mortality.

Globally, stroke incidence is rising, with early 3-month mortality reaching 15–30% due to individual and healthcare variations.^25,26 Stroke triggers multi-system pathophysiological changes, including neuroinflammation where neutrophils and platelets play pivotal roles.²⁷ In ischemic stroke, neutrophils rapidly infiltrate brain tissue via chemokine signaling (eg, CXCL1/CXCR2),²⁸ releasing ROS, proteases (MMP-9), and cytokines (IL-1β, TNF-α) that exacerbate neuronal and vascular injury.²⁹ Neutrophil-endothelial adhesion (CD11b/CD18-ICAM-1) further impairs microcirculation.³⁰ Platelet activation amplifies this process: neutrophil-platelet aggregates form through receptors (eg, PSGL-1/P-selectin, Mac-1/GPIIb/IIIa),³¹ driving (1) pro-thrombotic NETs release,³² (2) inflammatory amplification via RANTES/HMGB1,³³ and (3) thrombolytic resistance through PAI-1.³⁴ Targeting these interactions may mitigate secondary injury in LAO-AIS.

Research indicates that mechanical thrombectomy in stroke patients can further reactivate the inflammatory response, with neutrophil infiltration occurring as early as 180 minutes after ischemia and 30 minutes post-reperfusion.³⁵ Thus, neutrophils serve as the initiating factor in the inflammatory cascade, both following cerebral ischemia and after vessel recanalization via thrombectomy. The risk of reocclusion after thrombectomy is closely linked to platelet activity. Studies have demonstrated that thrombi retrieved during thrombectomy, particularly those rich in platelets and fibrin (such as arterial sources), exhibit poor responsiveness to both thrombolysis and mechanical thrombectomy, potentially increasing the likelihood of reocclusion.^36–38, Preoperative intravenous thrombolysis with recombinant tissue plasminogen activator (rtPA) may alter thrombus composition by degrading fibrin, thereby reducing platelet consumption after thrombectomy and potentially mitigating the risk of reocclusion.³⁷

Current research suggests that elevated levels of both neutrophils and platelets in stroke patients may contribute to poor prognosis. However, no well-established indicator specifically integrates these two parameters to assess early mortality after thrombectomy. Most existing studies rely on the platelet-to-lymphocyte ratio (PLR) and neutrophil-to-lymphocyte ratio (NLR) to predict adverse cardiovascular and cerebrovascular events,^39–41 while few directly link neutrophils and platelets. This may be due to the significant differences in their values and units, which could introduce errors in multiple regression and machine learning analyses. To accurately investigate the combined impact of neutrophils and platelets on early mortality in LAO-AIS, this study first applied centering and standardization to these values. While this approach may introduce some challenges in interpretation, it ensures the stability and balance of the model. The Boruta feature selection method was then employed to identify the most influential variables, effectively reducing the computational burden during model training and prediction while maintaining or even enhancing model performance.⁴² Boruta has been widely applied in predicting mortality risks across various diseases, including cardiovascular diseases, acute kidney injury, and sepsis-related acute kidney injury.^43–45 Multivariate regression analysis in this study identified NPR_std, Age, DC, Responsible arterie, LYM, and PT as independent predictors (P < 0.05), and a nomogram was developed accordingly. Prior studies have consistently highlighted age as a key factor influencing early mortality after thrombectomy in LAO-AIS patients, with significantly increased mortality rates observed in those over 80 years old.⁴⁶ Additionally, posterior circulation artery occlusion and prolonged PT have been identified as significant risk factors for mortality,^47–49 aligning with the findings of this study. Interestingly, SHAP analysis of PT revealed a biphasic effect, with both excessively short and prolonged PT values markedly influencing SHAP scores, suggesting a complex role of PT in stroke prognosis. Furthermore, timely DC has been shown to preserve brain function and significantly reduce mortality in patients with extensive cerebral infarction and severe cerebral edema.⁵⁰ Notably, this study diverged from previous research by identifying elevated LYM as a predictor of poor prognosis. Most studies have examined the prognostic implications of NLR or other blood cell ratios (eg, PLR, LMR) in acute ischemic stroke, often linking elevated NLR with adverse outcomes and suggesting that lymphopenia contributes to poor prognosis.^51–53 One study also reported that elevated total WBC counts, including lymphocytes, at admission were associated with increased three-month mortality following thrombolysis or thrombectomy but did not differentiate among lymphocyte subtypes.⁵⁴ These findings indicate that further preclinical research is needed to elucidate the precise mechanisms underlying the impact of lymphocyte levels on stroke prognosis.

It is noteworthy that although LYM and prothrombin time (PT) were identified as independent predictors in the multivariate regression model (Table 2), their individual discriminative capacity in univariate ROC analysis was limited (Figure 5). This apparent discrepancy is a recognized statistical phenomenon, wherein powerful predictors like NPR_std and Age may obscure the more subtle, yet independent, contributions of other variables in a univariate setting. Within the multivariate context, LYM and PT provide significant incremental value by capturing unique pathophysiological pathways—pertaining to immunomodulation and coagulation, respectively—that are not redundant with other model features. This justifies their inclusion in the final nomogram, which is designed as an integrated tool where the combined effect of all variables yields a robust prediction, as evidenced by the high overall AUC, rather than relying on any single parameter.

This study employed Boruta feature selection and applied centering and standardization to NPR, enhancing the stability and reliability of the model. The developed nomogram provides a practical tool for estimating early mortality risk in LAO-AIS patients following thrombectomy. Among various machine learning models evaluated, XGBoost was identified as the most suitable method based on multiple performance indicators. Furthermore, SHAP analysis highlighted the significance of NPR in the model, offering a clearer visualization of how different variable distributions influence SHAP values. This allows for a more intuitive understanding of the importance of each predictor at varying levels. Unlike previous studies that indirectly assess the roles of neutrophils and platelets through composite ratios (eg, NLR, PLR) or focus solely on LAO-AIS prognosis without considering thrombectomy, this study directly incorporates NPR as a predictive factor. By doing so, it provides a more clinically relevant approach, facilitating targeted interventions for patients post-thrombectomy—an aspect not addressed in prior research.

Beyond stroke-specific contexts, elevated neutrophil-to-platelet ratio (NPR) demonstrates significant clinical relevance across diverse conditions. In sepsis patients, high NPR correlates with adverse prognosis (eg, increased mortality), and multivariate analysis confirms its independent predictive value.⁵⁵ For burn patients, NPR combined with other inflammatory indices (eg, NLR, PLR) predicts in-hospital mortality, with the high-NPR group exhibiting significantly reduced survival rates.⁵⁶ Following cardiovascular surgery, NPR associates with perioperative complications (eg, infections, organ failure), where elevated NPR on the first postoperative day indicates poor outcomes.⁵⁷ In AIS patients, NPR links to hemorrhage risk after intravenous thrombolysis, and the high-NPR group shows higher adverse event incidence.⁵⁸ Furthermore, in testicular germ cell tumors (GCT), NPR correlates with tumor staging and clinical characteristics, with high-NPR patients demonstrating increased propensity for disease progression or metastasis.⁵⁹ These findings collectively underscore NPR’s role as a biomarker of inflammatory-thrombotic dysregulation.

To bridge our findings to bedside application, we propose a concrete clinical workflow for the proposed nomogram. Upon admission of an LAO-AIS patient eligible for MT, routine baseline data is collected. This includes a complete blood count (for calculating NPR_std and obtaining LYM) and a coagulation profile (for PT), alongside standard assessments such as neuroimaging (eg, CTA/DSA for determining the Resp. Art). NIHSS scoring, documentation of Age, and evaluation for potential DC. Risk Stratification: Prior to or immediately following MT, the clinician inputs these six parameters (NPR_std, Age, DC, Resp. Art., LYM, PT) into the nomogram to generate a quantifiable probability of 90-day mortality. This output stratifies patients into distinct risk categories (eg, low, intermediate, high). Personalized Clinical Decision-Making: This risk score is then integrated into clinical pathways to personalize post-procedural care: High-Risk Patients (eg, a hypothetical 75-year-old with basilar artery occlusion and elevated NPR_std): Would be flagged for intensified monitoring in a stroke unit or neuro-ICU. This facilitates preemptive management, such as stricter hemodynamic control to mitigate reperfusion injury, proactive screening and treatment of infections given the pro-inflammatory state, and early discussion with families regarding prognosis and goals of care. Low-Risk Patients: This prediction can aid in avoiding unnecessary overtreatment and rationalize resource allocation, allowing clinical focus to be directed toward higher-risk individuals. This practical framework demonstrates how our model directly addresses the stated research goal: to provide a clinically actionable tool that leverages NPR and other readily available parameters to guide post-thrombectomy management, thereby enhancing personalized medicine in LAO-AIS.

Notably, the socioeconomic context of Huizhou, Guangdong—where our cohort was recruited—offers valuable insights for researchers in Asian low- and middle-income countries (LMICs): as a transitional economy with a 2023 per capita GDP of ~$13,203 (surpassing typical LMIC thresholds) yet persistent urban-rural income gaps (2024 Q1-Q3: $5990 urban vs $3480 rural), it mirrors the economic duality of many upper-middle-income Asian regions. This context shaped our cohort’s key traits: high mechanical thrombectomy (MT) accessibility (evidenced by our 37.5% MT rate) stems from Huizhou’s robust stroke care infrastructure (eg, regional stroke centers, >95% health insurance coverage) and policy-driven investments (eg, “Sanming healthcare reform” funds), providing a replicable model for LMICs scaling MT. Conversely, our cohort’s 41.2% prevalence of uncontrolled hypertension reflects lingering gaps—such as low health literacy (34.22%)—that echo chronic disease management challenges in resource-constrained LMICs. Thus, our findings balance two critical takeaways for Asian LMICs: the feasibility of high-quality MT via targeted healthcare investment, and the need to address preventive care gaps to improve stroke outcomes.

Limitations

This study has several limitations that warrant emphasis. Its single-center, retrospective design inherently constrains the generalizability of our findings. The model was derived from a specific patient cohort treated under localized protocols, reflecting unique demographic and socioeconomic characteristics. Consequently, the predictive performance and clinical applicability of the nomogram may vary when applied in other geographic or healthcare settings. External validation through multi-center, prospective studies remains essential to confirm its broad utility.

Furthermore, several methodological constraints should be acknowledged. Key procedural variables—such as recanalization status and number of passes—as well as post-operative medications were not systematically analyzed, yet these may independently influence mortality risk. Although blood sampling within 24 hours captured acute inflammatory responses, the predictive stability of NPR could be affected by pre-analytical variations and the timing of measurement relative to thrombectomy. Future studies incorporating standardized pre- and post-procedural biomarker assessments are needed to enhance clinical translatability.

Conclusion

In conclusion, this study developed a predictive model for early mortality risk in LAO-AIS patients following thrombectomy. While the model incorporates six variables, its predictive capability is strongly driven by key factors, most notably the standardized Neutrophil-to-Platelet Ratio (NPR_std) and patient age, with SHAP analysis visually confirming the pivotal role of NPR_std in the decision-making process.

Funding Statement

This research was supported by the Major Project (2023Y9316) of the Fujian Provincial Department of Science and Technology.

Abbreviation

NPR, neutrophil-to-platelet ratio; MT, mechanical thrombectomy; LAO-AIS, large artery occlusive acute ischemic stroke; DT, decision tree; XGBoost, extreme gradient boosting; SVM, support vector machine; NB, naive bayes; NPR_std, centralized and standardized neutrophil-to-platelet ratio; DC, decompressive craniectomy; LYM, lymphocyte count; PT, prothrombin time; C-index, concordance index; DCA, clinical decision curve analysis; AUC, area under the ROC curve; F1, F1-score; SHAP, shapley additive explanations; DALYs, disability-adjusted life years; LAO, large artery occlusion; AIS, acute ischemic stroke; LMICs, low- and middle-income countries; CTA, CT angiography; DSA, digital subtraction angiography; ASPECTS, alberta stroke program early CT score; MRI, magnetic resonance imaging; DAWN, deferring antiplatelet therapy until symptomatic intracranial stenosis; DEFUSE-3, diffusion and perfusion imaging evaluation for understanding stroke evolution with intravenous alteplase; BAOCHE trial, basilar artery Occlusion chinese endovascular trial; LDL-C, low-density lipoprotein cholesterol; HDL-C, high-density lipoprotein cholesterol; AIC, akaike information criterion; BIC, bayesian information criterion; ROC, receiver operating characteristic; SBP, systolic blood pressure; DBP, diastolic blood pressure; WBC, white blood cell; NE, neutrophil granulocyte; PLT, platelet; MO, monocyte; APTT, activated partial thromboplastin time; D-D, d-dimer; FIB, fibrinogen; ALB, albumin; K, potassium; Ca, calcium; Cl, chlorine; PC, posterior circulation; AC, anterior circulation; Resp. Art., responsible artery; LR, logistic regression; Dfclass, different classes; rtPA, recombinant tissue plasminogen activator; PLR, platelet-to-lymphocyte ratio; NLR, neutrophil-to-lymphocyte ratio; GCT, germ cell tumors; LMICs, low- and middle-income countries.

Data Sharing Statement

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding authors.

Ethical Statement

The study involving human participants was reviewed and approved by the Ethics Committee of the Affiliated Hospital of Guilin Medical University (Approval No. 2024YJSLL-18). The committee waived the requirement for written informed consent due to the retrospective nature of the study and in accordance with national legislation and institutional policies.

Author Contributions

Shaohuai Xia: Writing – original draft; Formal analysis; Data curation; Methodology. Junhong Hu: Writing – original draft; Data curation; Formal analysis. Yingye Liao: Writing – original draft; Investigation (Data collection). Jing Wu: Writing – original draft; Investigation (Data collection). Xinrong Zhong: Writing - original draft, Investigation (Data collection). Xiaoguang Fan: Writing – original draft; Formal analysis; Investigation. Jinping Li: Writing – review and editing, Investigation. Guifeng Liang: Writing – review and editing, Investigation. Li Chen: Conceptualization; Supervision; Writing – review and editing. Xuewei Xia: Conceptualization; Supervision; Writing – review and editing. All authors have read and approved the final manuscript. All authors have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare no conflict of interest.